# I Built a Local, GPU-Accelerated Voice Commander—And I Still Type Everything > Source: > Published: 2026-06-16 18:12:29+00:00 As developers, we love building productivity tools. We spend days optimizing our development environments, writing shell scripts, configuring keybindings, and building custom pipelines. A few months ago, I built [Voice Commander](https://github.com/MasihMoafi/Voice-commander)—a local, GPU-accelerated voice transcription system. It uses a local CUDA-powered **Whisper.cpp** model to transcribe my speech, and hooks into the **Gemini API** to clean up filler words ("um", "uh"), fix grammar, structure the output, and auto-paste it directly at my cursor. It works like a charm. It is fast, private, and precise. And yet, **I still type everything.** Every time I need to write a long code comment, draft an issue, or even outline a plan, my hands immediately go to the keyboard. I catch myself typing away, while the microphone hotkey is right there, ready to save me hundreds of keystrokes. Why is this? And why is it so hard to break the keyboard habit? Through this, I realized that typing isn't just a physical action; it's a cognitive buffer. Refusing to use a tool you built because it "feels weird" is a common trap. To break this, I'm forcing myself to follow a new rule: **if a thought requires more than two sentences, I must dictate it.** By pushing through the initial awkwardness, I'm hoping to make voice-to-text as natural as reaching for the trackpad. Have you built tools that you struggle to integrate into your actual daily routine? How do you overcome the muscle memory of typing? *If you want to try out the project locally, check it out on GitHub: **[MasihMoafi/Voice-commander](https://github.com/MasihMoafi/Voice-commander)*** *For more of my AI research and developer tools, visit my website: **[masihmoafi.tech](https://masihmoafi.tech)***